RTX 5090 · Qwen-Image-2.1 · 2048×2048 · same prompt, same seed

29 s instead of 108 s.

The new Turbo model needs 8 steps. The base model needs 40. Here are all 6 test images.

Base · 40 steps

108 s

Turbo · 8 steps

29 s

⇆Drag across each image: Base on the left, Turbo on the right.

Setup and limits

Times are end-to-end per image, prompt in to PNG out, median over the 6 prompts. Turbo: 1 warm-up, then 3 runs per prompt. Base: 1 run per prompt, no warm-up.
Poster text: 11 of 11 words right in both. Small UI text is garbled in both; only “18°” reads correctly. Image quality here is my eye, not a metric.
RTX 5090 at 450 W, driver 610.43.02, torch 2.11 + cu128, diffusers from git (1d5d056), transformers 5.19. Both models with model CPU offload (32.4 GB of bf16 weights on a 32 GB card), peak GPU memory 27,570 MiB.
Models: Qwen/Qwen-Image-2.1-Turbo @0da70ce (released 9 October 2026), Qwen/Qwen-Image-2.1 @d26bb61.